Automated Computer-Aided Diagnosis of Diabetic Retinopathy Based on Segmentation and Classification using K-nearest neighbor algorithm in retinal images

Abstract

Abstract Diabetic retinopathy (DR) is an eye oddity where the human retina is afflicted because of the ever-increasing quantity of insulin in the blood. It leads to the loss of sight. Preliminary diagnosis of DR assists to improve to inhibit future injury. Proper DR screening has been recognized as an economical way to accumulate health services. Automated retinal analysis become known as the most significant screening approach for primitive DR diagnosis, which leads to diminishing the workload related to manual screening and also, cost-effective and less time-consuming efforts. In the proposed work, the preprocessing, removal of applicant lesion pixels, and formulation of feature set have been examined which is fully appropriate for the classification task. In preprocessing approach, the framework removes the unwanted pixels, eliminates the optic disc, and extraction of the blood vessels from the retinal images. Morphological operations are applied to extract the boundaries of the blood vessels and then 2D discrete wavelet decomposition is applied to estimate the horizontal, vertical and diagonal coefficients. The candidate lesion pixels i.e. dark and bright DR pixels are detected using an adaptive threshold that uses local statistical, geometrical, and location-based characteristics of the background image. The extracted feature set is processed using a K-nearest neighbor (KNN) classifier with 80% of training data and 20% of testing data to diagnose the severity level of the disease. The proposed scheme is evaluated by the DIARETDB1 benchmark dataset with the performance parameters, i.e. 95% of accuracy, 92.6% of sensitivity and 87.56% specificity achieved with less computation time required.